Co-Clustering by Bipartite Spectral Graph Partitioning for Out-of-Tutor Prediction

Learning a more distributed representation of the input feature space is a powerful method to boost the performance of a given predictor. Often this is accomplished by partitioning the data into homogeneous groups by clustering so that separate models could be trained on each cluster. Intuitively each such predictor is a better representative of the members of the given cluster than a predictor trained on the entire data-set. Previous work has used this basic premise to construct a simple yet strong bagging strategy. However, such models have one significant drawback: Instances (such as students) are clustered while features (tutor usage features/items) are left alone. One-way clustering by using some objective function measures the degree of homogeneity between data instances. Often it is noticed that features also influence final prediction in homogeneous groups. This indicates a duality in the relationship between clusters of instances and clusters of features. Co-Clustering simultaneously measures the degree of homogeneity in both data instances and features, thus also achieving clustering and dimensionality reduction simultaneously. Students and features could be modelled as a bipartite graph and a simultaneous clustering could be posed as a bipartite graph partitioning problem. In this paper we integrate an effective bagging strategy with Co-Clustering and present results for prediction of out-of-tutor performance of students. We report that such a strategy is very useful and intuitive, even improving upon performance achieved by previous work.